LIBRO DE ACTAS (pdf) - Universidad de Sevilla

Automatic Detection of Melodic Patterns in Flamenco Singing by Analyzing Polyphonic Music Recordings
perceive as the patterns appears in different modal contexts. When the similarity threshold is
set to 70 % and 60 %, the algorithm still finds the pattern correctly in the Huelva capital
corpus. We can state that pattern is there, more or less blurry or sketched, but still present.
In the Valverde corpus, at 80 % the algorithm finds this pattern only in the cantes sung by girls
trained at flamenco clubs in Huelva. Those clubs are called peñas flamencas. Normally, peñas
organize singing lessons. Girls from peñas are trained to sing very standard models and,
therefore, they do not contribute in terms of innovation as other singers of fandango do (for
instance, singers such as Toronjo or Rengel).
At 70 % and below this threshold, we find the well-established, acclaimed voices of fandango
singing. The similarity distance increases on Toronjo, perhaps due to its particular vocal
technique. Raya, another great figure, is also detected by the algorithm. Here we can pose the
question of whether the system is sensitive to the hoarse voice of singers like Toronjo as
opposed to the clean voices of the girls of flamenco clubs. Furthermore, the algorithm seems
to be sensitive to dynamic. Toronjo has peaks and valleys in dynamics, where the children
school has a rather flat singing, much less expressive. The algorithm detects the canon but not
the personal “trait".
• Exp-6: This is a pattern used as preparation for the final cadence of the last phrase. In the
Huelva capital corpus and at 60% and 70% of the similarity value, the algorithm finds it at a
macro-structure level, that is, it finds it in the beginning, in the middle, and in the final section.
When the similarity threshold is raised to 80%, then it is only found in the final cadence. For
the Valverde corpus, at all levels of similarity, the algorithm returns correct results, although as
pointed out above, it still misses many results. Again, at high similarity values, the pattern is
only found in the final cadence.
6. Conclusions
In this work we have carried out a study of fandango styles through the analysis of archetypal
melodic patterns. The problem that we tackled comprised different aspects. As already insisted,
in flamenco no written scores are in general available; hence, symbolic analysis have left for
future research. We designed an algorithm to extract a MIDI-like representation from the
fandango corpus that at the same time were meaningful and tractable. After this step, a algorithm
was designed to find the patterns in the given corpora. Our algorithm has proved to be robust
inasmuch it was able to find the abstract patterns in the corpora in spite of the fact that music
was polyphonic, there was a great deal of melismas as well as a high degree of tempo deviation.
Still, some computational aspects of the algorithm could be still improved, such as time
complexity. On the musicological side, we investigated the problem of finding certain archetypal
patterns specified by flamenco experts in actual corpora of fandango styles. The presence and
the position of the melodic patterns in the fandango provides important information. For
instance, it could help to understand the evolution of fandango styles. Also, a complementary
approach to analysing fandango style is Schenkerian analysis (see [Esc12]). This kind of analysis
would provide a better understanding of the deep structure of these styles.
As for future work, this study could be extended to other Huelva fandango styles. A more
ambitious goal would be to carry out the analysis for the whole corpus of fandango. Also, other
musical features could be taken into account and thus perform a more general analysis, not only
based on melody. In particular, form or stylistic ornamentations are suitable candidates for that
potential analysis. Regarding the pattern detection algorithm, the main improvement lies in the
number of returned results over the number of total expected results (see Table 1), which is still
too low. The audio feature extraction algorithm can also be refined, specially in the source
separation step.
Acknowledgments
This research is partially supported by project COFLA: Computational Analysis of Flamenco
232